Margin-Based Deep Learning Networks for Human Activity Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Human Activity Recognition (HAR) is crucial for many applications.
• Deep learning models show promise but struggle with similar activities and inter-personal variability.
• Existing HAR methods face challenges in intra-class scatter and inter-class separation.

Purpose of the Study:

• To enhance the discriminative power of deep learning networks for HAR.
• To address the limitations of recognizing similar activities and activities with high user variability.
• To evaluate a novel margin mechanism for improved HAR performance.

Main Methods:

• Introduced a margin mechanism to modify common neural networks.
• Applied the margin-based models to four different neural network architectures.
• Evaluated performance on the OPPORTUNITY, UniMiB-SHAR, and PAMAP2 datasets.
• Extended the research to open-set human activity recognition.

Main Results:

• Margin-based models demonstrated superior performance compared to unmodified models.
• The proposed method effectively improved the classification of human activities.
• Successful evaluation in recognizing new human activities within an open-set framework.

Conclusions:

• The margin mechanism is effective in enhancing deep learning for HAR.
• The approach successfully tackles challenges of activity similarity and user variability.
• The method shows potential for real-world HAR applications and open-set recognition.